1 / 67

A Framework for Asynchronous Parallel Machine Learning

A Framework for Asynchronous Parallel Machine Learning. Joseph Gonzalez Joint work with. Yucheng Low. Aapo Kyrola. Danny Bickson. Carlos Guestrin. Guy Blelloch. Joe Hellerstein. David O’Hallaron. Alex Smola. How will we design and implement parallel learning systems?.

adler
Download Presentation

A Framework for Asynchronous Parallel Machine Learning

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. A Framework for Asynchronous Parallel Machine Learning Joseph Gonzalez Joint work with Yucheng Low Aapo Kyrola Danny Bickson Carlos Guestrin Guy Blelloch Joe Hellerstein David O’Hallaron Alex Smola

  2. How will wedesign and implementparallel learning systems?

  3. We could use …. Threads, Locks, & Messages “low level parallel primitives”

  4. Threads, Locks, and Messages • ML experts repeatedly solve the same parallel design challenges: • Implement and debug complex parallel system • Tune for a specific parallel platform • Two months later the conference paper contains: “We implemented ______ in parallel.” • The resulting code: • is difficult to maintain • is difficult to extend • couples learning model to parallel implementation Graduatestudents

  5. ... a better answer: Map-Reduce / Hadoop Build learning algorithms on-top of high-level parallel abstractions

  6. MapReduce – Map Phase 4 2 . 3 2 1 . 3 2 5 . 8 CPU 1 1 2 . 9 CPU 2 CPU 3 CPU 4 Embarrassingly Parallel independent computation No Communication needed

  7. MapReduce – Map Phase 8 4 . 3 1 8 . 4 8 4 . 4 CPU 1 2 4 . 1 CPU 2 CPU 3 CPU 4 1 2 . 9 4 2 . 3 2 1 . 3 2 5 . 8 Image Features

  8. MapReduce – Map Phase 6 7 . 5 1 4 . 9 3 4 . 3 CPU 1 1 7 . 5 CPU 2 CPU 3 CPU 4 8 4 . 3 1 8 . 4 8 4 . 4 1 2 . 9 2 4 . 1 4 2 . 3 2 1 . 3 2 5 . 8 Embarrassingly Parallel independent computation No Communication needed

  9. MapReduce – Reduce Phase Attractive Face Statistics Ugly Face Statistics 17 26 . 31 22 26 . 26 CPU 1 CPU 2 1 2 . 9 2 4 . 1 1 7 . 5 4 2 . 3 8 4 . 3 6 7 . 5 2 1 . 3 1 8 . 4 1 4 . 9 2 5 . 8 8 4 . 4 3 4 . 3 Image Features

  10. Map-Reduce for Data-Parallel ML • Excellent for large data-parallel tasks! Data-ParallelGraph-Parallel Is there more to Machine Learning ? Map Reduce Label Propagation Lasso Feature Extraction Cross Validation Belief Propagation Kernel Methods Computing Sufficient Statistics Tensor Factorization PageRank Neural Networks Deep Belief Networks

  11. Concrete Example Label Propagation

  12. Label Propagation Algorithm • Social Arithmetic: • Recurrence Algorithm: • iterate until convergence • Parallelism: • Compute all Likes[i] in parallel Sue Ann 50% What I list on my profile 40% Sue Ann Likes 10% Carlos Like 80% Cameras 20% Biking 40% + I Like: 60% Cameras, 40% Biking Profile 50% 50% Cameras 50% Biking Me Carlos 30% Cameras 70% Biking 10%

  13. Properties of Graph Parallel Algorithms Dependency Graph Factored Computation Iterative Computation What I Like What My Friends Like

  14. Map-Reduce for Data-Parallel ML • Excellent for large data-parallel tasks! Data-ParallelGraph-Parallel Map Reduce Map Reduce? ? Label Propagation Lasso Feature Extraction Cross Validation Belief Propagation Kernel Methods Computing Sufficient Statistics Tensor Factorization PageRank Neural Networks Deep Belief Networks

  15. Why not use Map-Reducefor Graph Parallel Algorithms?

  16. Data Dependencies • Map-Reduce does not efficiently express dependent data • User must code substantial data transformations • Costly data replication Independent Data Rows

  17. Iterative Algorithms • Map-Reduce not efficiently express iterative algorithms: Iterations Data Data Data Data CPU 1 CPU 1 CPU 1 Data Data Data Data Data Data Data Data CPU 2 CPU 2 CPU 2 Data Data Data Data Data Data Data Data CPU 3 CPU 3 CPU 3 Data Data Data Slow Processor Data Data Data Data Data Barrier Barrier Barrier

  18. MapAbuse: Iterative MapReduce • Only a subset of data needs computation: Iterations Data Data Data Data CPU 1 CPU 1 CPU 1 Data Data Data Data Data Data Data Data CPU 2 CPU 2 CPU 2 Data Data Data Data Data Data Data Data CPU 3 CPU 3 CPU 3 Data Data Data Data Data Data Data Data Barrier Barrier Barrier

  19. MapAbuse: Iterative MapReduce • System is not optimized for iteration: Iterations Data Data Data Data CPU 1 CPU 1 CPU 1 Data Data Data Data Data Data Data Data CPU 2 CPU 2 CPU 2 Data Data Data StartupPenalty Disk Penalty Disk Penalty Startup Penalty Startup Penalty Disk Penalty Data Data Data Data Data CPU 3 CPU 3 CPU 3 Data Data Data Data Data Data Data Data

  20. Map-Reduce for Data-Parallel ML • Excellent for large data-parallel tasks! Data-ParallelGraph-Parallel Map Reduce Pregel (Giraph)? Map Reduce? SVM Lasso Feature Extraction Cross Validation Belief Propagation Kernel Methods Computing Sufficient Statistics Tensor Factorization PageRank Neural Networks Deep Belief Networks

  21. Pregel (Giraph) • Bulk Synchronous Parallel Model: Compute Communicate Barrier

  22. Problem Bulk synchronous computation can be highly inefficient. Example:Loopy Belief Propagation

  23. Loopy Belief Propagation (Loopy BP) • Iteratively estimate the “beliefs” about vertices • Read in messages • Updates marginalestimate (belief) • Send updated out messages • Repeat for all variablesuntil convergence

  24. Bulk Synchronous Loopy BP • Often considered embarrassingly parallel • Associate processor with each vertex • Receive all messages • Update all beliefs • Send all messages • Proposed by: • Brunton et al. CRV’06 • Mendiburu et al. GECC’07 • Kang,et al. LDMTA’10 • …

  25. Sequential Computational Structure

  26. Hidden Sequential Structure

  27. Hidden Sequential Structure • Running Time: Evidence Evidence Time for a single parallel iteration Number of Iterations

  28. Optimal Sequential Algorithm Running Time Bulk Synchronous 2n2/p Gap Forward-Backward 2n p ≤ 2n p = 1 n Optimal Parallel p = 2

  29. The Splash Operation • Generalize the optimal chain algorithm:to arbitrary cyclic graphs: ~ Grow a BFS Spanning tree with fixed size Forward Pass computing all messages at each vertex Backward Pass computing all messages at each vertex

  30. Data-Parallel Algorithms can be Inefficient Optimized in Memory Bulk Synchronous Asynchronous Splash BP The limitations of the Map-Reduce abstraction can lead to inefficient parallel algorithms.

  31. The Need for a New Abstraction • Map-Reduce is not well suited for Graph-Parallelism Data-ParallelGraph-Parallel Map Reduce Pregel (Giraph) Feature Extraction Cross Validation Belief Propagation Kernel Methods SVM Computing Sufficient Statistics Tensor Factorization PageRank Lasso Neural Networks Deep Belief Networks

  32. What is GraphLab?

  33. The GraphLab Framework Scheduler Graph Based Data Representation Update Functions User Computation Consistency Model

  34. Data Graph A graph with arbitrary data (C++ Objects) associated with each vertex and edge. • Graph: • Social Network • Vertex Data: • User profile text • Current interests estimates • Edge Data: • Similarity weights

  35. Implementing the Data Graph Multicore Setting Cluster Setting In Memory Partition Graph: ParMETIS or Random Cuts Cached Ghosting • In Memory • Relatively Straight Forward • vertex_data(vid)  data • edge_data(vid,vid)  data • neighbors(vid)  vid_list • Challenge: • Fast lookup, low overhead • Solution: • Dense data-structures • Fixed Vdata& Edata types • Immutable graph structure A B C D Node 1 Node 2 A B A B C D C D

  36. The GraphLab Framework Scheduler Graph Based Data Representation Update Functions User Computation Consistency Model

  37. Update Functions An update function is a user defined program which when applied to a vertex transforms the data in the scopeof the vertex label_prop(i, scope){ // Get Neighborhood data (Likes[i], Wij, Likes[j]) scope; // Update the vertex data // Reschedule Neighbors if needed if Likes[i] changes then reschedule_neighbors_of(i); }

  38. The GraphLab Framework Scheduler Graph Based Data Representation Update Functions User Computation Consistency Model

  39. The Scheduler The scheduler determines the order that vertices are updated. b d a c CPU 1 c b e f g Scheduler e f b a i k h j i h i j CPU 2 The process repeats until the scheduler is empty.

  40. Implementing the Schedulers Multicore Setting Cluster Setting Multicore scheduler on each node Schedules only “local” vertices Exchange update functions • Challenging! • Fine-grained locking • Atomic operations • Approximate FiFo/Priority • Random placement • Work stealing Node 1 Node 2 CPU 1 CPU 1 CPU 2 CPU 2 f(v1) CPU 1 CPU 2 CPU 3 CPU 4 Queue 1 Queue 1 Queue 2 Queue 2 Queue 1 Queue 2 Queue 3 Queue 4 f(v2) v1 v2

  41. The GraphLab Framework Scheduler Graph Based Data Representation Update Functions User Computation Consistency Model

  42. GraphLab Ensures Sequential Consistency For each parallel execution, there exists a sequential execution of update functions which produces the same result. time CPU 1 Parallel CPU 2 Single CPU Sequential

  43. Ensuring Race-Free Code • How much can computation overlap?

  44. Consistency Rules Full Consistency Data Guaranteed sequential consistency for all update functions

  45. Full Consistency Full Consistency

  46. Obtaining More Parallelism Full Consistency Edge Consistency

  47. Edge Consistency Edge Consistency Safe Read CPU 1 CPU 2

  48. Consistency Through R/W Locks • Read/Write locks: • Full Consistency • Edge Consistency Write Write Write Canonical Lock Ordering Read Read Write Read Write

  49. Consistency Through R/W Locks • Multicore Setting: Pthread R/W Locks • Distributed Setting: Distributed Locking • Prefetch Locks and Data • Allow computation to proceed while locks/data are requested. Node 1 Node 2 Data Graph Partition Lock Pipeline

  50. Consistency Through Scheduling • Edge Consistency Model: • Two vertices can be Updated simultaneously if they do not share an edge. • Graph Coloring: • Two vertices can be assigned the same color if they do not share an edge. Phase 1 Phase 2 Phase 3 Barrier Barrier Barrier

More Related